菜单
ROCm profiling workflow for AMD GPU kernels using rocprofv3 and rocprof-compute. Use when profiling hot kernels, collecting counters, diagnosing memory-vs-compute-vs-stall bottlenecks, reading Perfetto traces, or validating low-precision AMD kernels.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | tool-rocprof |
| description | ROCm profiling workflow for AMD GPU kernels using rocprofv3 and rocprof-compute. Use when profiling hot kernels, collecting counters, diagnosing memory-vs-compute-vs-stall bottlenecks, reading Perfetto traces, or validating low-precision AMD kernels. |
Use this skill when you need runtime evidence before changing an AMD kernel.
| Tool | Best for | Output |
|---|---|---|
rocprofv3 | Timeline traces, raw counters, API/kernel overlap, Perfetto | CSV / JSON / PFTrace / rocpd |
rocprof-compute | Speed-of-Light, roofline, occupancy, MFMA / memory panels | Workload directory + CLI / GUI analysis |
rocm-smi, rocminfo, hipcc --resource-usage | Environment, clocks, device properties, register / spill clues | Text diagnostics |
Rule of thumb:
rocprof-compute if you want a fast high-level bottleneck classification.rocprofv3 when you need exact trace timing, kernel overlap, or raw counters.rocprof-compute on that focused benchmark to get Speed-of-Light and occupancy context.rocprofv3 for timeline gaps, counter detail, or multi-kernel overlap.# High-level metric collection
rocprof-compute profile --name workload_name -- python run_kernel.py
rocprof-compute analyze -p workload_name/<GPU_NAME> --cli # <GPU_NAME> subdir is auto-named by the detected GPU, e.g. MI300X / MI350X
# Discover available metric blocks / help
rocprof-compute profile -h
rocprof-compute profile --list-metrics
rocprof-compute analyze -h
# Trace or raw-counter collection
rocprofv3 --help
rocprofv3 --list-counters
rocprofv3 <trace-or-counter-options> -- python run_kernel.py
# Supporting commands
rocm-smi
rocminfo
hipcc --resource-usage kernel.hip
The exact rocprofv3 flags vary by ROCm release, so verify them with local --help before scripting.
Use Speed-of-Light or equivalent GPU activity to decide whether microarchitecture tuning is even the right problem:
| Signal | Interpretation | Next step |
|---|---|---|
GPU utilization > 60% | GPU-bound enough to optimize at kernel level | Continue to memory vs compute split |
GPU utilization < 30% | Likely launch, CPU, synchronization, or scheduling issue | Inspect timeline / host-side gaps first |
30-60% | Mixed or unclear | Cross-check trace overlap and SoL panels |
Do not start with micro-tuning if the GPU is mostly idle.
Typical actions:
Typical actions:
Typical actions:
These are starting heuristics, not laws:
| Metric | Good | Needs attention |
|---|---|---|
| Occupancy | > 50% | < 25% |
| HBM utilization | > 60% when memory-bound | < 30% |
| MFMA utilization | > 50-70% when compute-bound | < 40% |
| LDS bank conflict ratio | < 5% | >= 5% |
| Scratch / spills | 0 | Any non-zero |
Always compare against the target GPU's known good/bad ranges for gfx942 / gfx950: use the project knowledge base consulted during ANALYZE plus the agent's own hardware knowledge.
| Area | Typical counters / panels | Why |
|---|---|---|
| MFMA | SQ_INSTS_MFMA, SQ_INSTS_VALU_MFMA_MOPS_*, MFMA busy panels | Detect matrix-core usage and efficiency |
| VMEM / LDS | SQ_INSTS_VMEM_*, SQ_INSTS_LDS, SQ_LDS_BANK_CONFLICT | Detect bandwidth waste and LDS issues |
| Cache | TCC_HIT, TCC_MISS, TCP / TCC panels | Estimate reuse and coalescing quality |
| Occupancy / resources | Occupancy report, scratch metrics, hipcc --resource-usage | Find register or LDS pressure |
| Timeline | PFTrace / Perfetto | Detect launch gaps, overlap issues, sequencing problems |
Counter names vary across ROCm releases. Use local tool output to confirm the exact spelling.
Profiling is not enough for low-precision work. Every FP8 / FP6 / FP4 / MXFP round should explicitly record:
Critical rule:
gfx942 FP8 is typically FNUZgfx950 FP8 is OCPNever label a result only as "fp8" when multiple encodings are possible.
Validation strategy:
../../../../rules/iteration_rules.mdc../../../kernel-optimize/workflow/optimize-loop.md../../SKILL.md (Primus-Turbo development guide)Primus-Turbo handoff to the autonomous kernel-optimize loop — collect the prerequisites (kernel path, focused test/bench commands, scoring metric, execution mode, quick-validation harness) a kernel campaign needs and pass them on. Use when pushing a Primus-Turbo kernel toward the hardware limit, not just spot-checking perf.
Develop, debug, and validate Primus-Turbo operators and modules on AMD GPUs. Covers the layered architecture (ops / kernels-dispatcher / Triton / HIP-CK csrc / modules), how to add or change a feature end-to-end, accuracy verification (SNR, tolerances, reference implementations), performance benchmarking, the backend dispatch system, and build/test/bench commands. Use for any Primus-Turbo development task (GEMM, Attention, GroupedGEMM, MoE, quantization, normalization, activation) and for accuracy or performance validation.
AI-driven operator performance optimization framework. Defines the optimization loop, execution environment selection, knowledge routing, and logging conventions to drive agent-autonomous iteration toward hardware limits.
Primus-Turbo feature development workflow — the layered architecture (ops / kernels-dispatcher / Triton / HIP-CK csrc / modules), how to wire a new operator end-to-end, which layer to touch, and which existing file to copy. Use when adding or changing a Primus-Turbo operator or module on AMD GPUs.
Primus-Turbo performance verification — run single-operator and suite benchmarks, read the latency/TFLOPS metrics, source real-model shapes, and derive a combined training-step metric. Use when measuring latency or throughput of a Primus-Turbo operator.
Primus-Turbo accuracy verification — compare an operator against a higher-precision reference for forward and backward, with the right gate (allclose for bf16/fp16/fp32, SNR for fp8/fp4) and FP8 encoding awareness. Use when validating numerical correctness of a Primus-Turbo operator.